AI chip design startup Ricursive Intelligence has released Ripple-65B Preview, an open-source 65-billion parameter language model published under the Apache 2.0 license on Hugging Face. Pretrained on 12 trillion tokens, the model scored 80.52 on the TerminalBench 2.1 coding benchmark, which the company claims is the highest result among open US models.
According to Ricursive Intelligence, total research, development, and training costs for Ripple-65B Preview came out to $32 million. The model employs a dense Transformer architecture optimized specifically for Nvidia hardware, bypassing the mixture-of-experts design common in recent models. While benchmark results indicate it does not match the newly released Inkling 2 Small, the company notes that Ripple is easier for organizations to deploy on local hardware.
The model utilizes a harness combining Terminus 2 and OpenCode to achieve its coding task performance, targeting developers looking for open-weights alternatives for software engineering tasks.
Why it matters
Offers engineering teams a fully open Apache 2.0 65B parameter model optimized for coding and local Nvidia deployment.
Demonstrates that competitive dense coding models can be developed and trained for a relatively low R&D budget of $32 million.
Provides an open-weights alternative for enterprises seeking to avoid proprietary agentic coding API lock-in.
Source: the-decoder.com



